Just a quick summary for those who were following this issue, the real issue came from the fact that two distinct sockets were used by the script, one to send the command an another one to retrieve the stats. Since there is no guarantee on the order of delivery on two parallel sockets, it explains why it worked sometimes only, and why adding a delay made the process more reliable. Michele switched to a single socket which fixed the issue. Another solution is to wait for haproxy to ACK the command on the socket, but it's not necessarily always worth it.
I understood the single socket was Michele's original design but it was probably failing due to one bug recently fixed (tracking servers exiting maintenance mode). Hoping this will save troubleshooting time for other people. Regards, Willy
